Armory Agent Editors

Agents inherit permissions without checking if each task was actually authorized.

Attackers hide malicious instructions in tool descriptions that AI agents trust without question.

Malicious instructions spread through agent pipelines, gaining power at each stage.

Agents leak data through message chains because tool-level security misses compositional risk.

Persistent memory turns agents into targets for poisoning attacks that bypass traditional security.

Agents need validation at every data entry point, not just the initial user request.